M1: Add Ollama chat loop with LLM provider abstraction

[bug-ops]

Summary

• LlmProvider trait with native async (no async-trait crate, Edition 2024)
• OllamaProvider implementation using ollama-rs
• Config loading from TOML with env var overrides (ZEPH_LLM_PROVIDER, ZEPH_LLM_BASE_URL, ZEPH_LLM_MODEL)
• Interactive CLI agent loop with multi-turn conversation history
• Generic Agent<P: LlmProvider> for dyn-safety with native async traits
• 6 unit tests (config: 3, ollama: 3)

Acceptance Criteria

• cargo run starts interactive REPL
• Config loaded from config/default.toml
• Env vars override config values
• Multi-turn conversation within session
• Ctrl+C / "exit" / "quit" exits cleanly
• cargo clippy -- -D warnings clean
• cargo test --workspace — 6 tests pass

Closes #8, closes #9, closes #10, closes #11
Relates to #2